Distributed framework ray - detailed introduction to starting ray and connecting clusters

Ray The program can use the underlying Ray Parallelize and distribute the runtime .Ray The runtime consists of multiple services that are started in the background / Process composition , For communication 、 The data transfer 、 Scheduling, etc .Ray It can run on a laptop 、 Start on a single server or multiple servers .

start-up Ray There are three ways to run ：

1、 Implicit pass ray.init()（ Start on a single machine Ray, Automatically create ray colony ）
2、 adopt CLI Explicit （ adopt CLI start-up Ray (ray start --head --port=6379)）
3、 Explicitly through the cluster initiator （ start-up Ray colony (ray up ./xx.sh)）
4、 Use Ray The client connects to an existing Ray Runtime

Check ray Whether to initialize

import ray
ray.init()
#  Check ray Whether to initialize 
if ray.is_initialized() == False
	ray.init()
#  Show off api 
ray.shutdown()

ray client

Ray Client It is a will. Python Script connects to remote Ray Clustered API. actually , It allows you to run as if on a local computer Ray Also use remote Ray colony .
Use ray Client scenario ： Requires script or interactive connection to remote , Apply to ray1.5+

ray.init(
    "ray://1.2.3.4:10001",
    namespace="my_namespace",#  Set the namespace of the session 
    runtime_env={
    "working_dir": "files/my_project", "pip": ["toolz", "requests"]},
    #  Set the running environment of the session 
)

Ray Connected to the cluster 3 In the form of

For the convenience of introducing , Suppose the machine A,B,C
1、 if B and C Used to set up the cluster ,A Do development ,A Script to connect to the cluster , Need to use （ Client connection ）ray.init(“ray://<head_node_host>：10001”)
2、 if B Is the head node of the cluster , And in B On the development ,B Script to connect to the cluster , Need to use （ Non client connection ）
ray.init(“localhost:”)
3、 There is only one machine A, Automatically create clusters , Need to be （ Non client connection ） Use ray.init()